可靠性管理公司 Gremlin 宣布其智能体产品 Foresight AI 正式全面上线。该产品可分析服务中的潜在故障、给出优化建议，并重新执行测试以验证修复方案是否有效。它作为 Gremlin 平台的附加组件提供。该平台支持团队开展计划内实验、控制故障的影响范围，并使用自动化测试确保修复措施长期有效。企业可利用可靠性评分量化改进进度，优先投入资源提升系统弹性。
Foresight AI 的目标是提供可靠性检测能力，无需每位开发者都成为分布式系统专家。Gremlin 将这款产品定位为应对 AI 辅助软件快速交付的解决方案。Gremlin 首席执行官兼联合创始人 Kolton Andrus 认为，当前有大量代码在缺少人工深度审查的情况下直接进入了生产环境。
“最好的事故，是从未发生的事故。”
该产品将工作划分为四个智能体角色：分析师智能体（Analyst）收集服务数据并推荐测试用例，测试员智能体（Tester）负责安排和执行测试，运维智能体（Operator）解读失败的测试结果并提出修复方案，技术项目经理智能体（Technical Program Manager）跟踪各团队的测试覆盖率、可靠性风险和承诺事项。最后一个智能体可以生成每周报告，并通过 Slack 推送更新。
Foresight AI 的推荐基于 Gremlin 的故障图谱（Failure Atlas）。该公司称，这是一个专有的记录库，包含过去十多年在数万个分布式系统上进行的数百万次故障注入实验。Gremlin 表示，系统数据不会被用于训练大语言模型。智能体使用平台数据和故障图谱进行决策，而 LLM 仅用于搜索、总结和解释结果。
其主要技术特点是闭环验证机制。当测试发现系统薄弱点时，服务会提出变更建议，然后重新运行原始测试。Gremlin 表示，Foresight AI 在未获得许可的前提下不会执行任何测试或应用修复方案。这样既保持由工程师对变更负责，又把大部分分析和重复验证工作自动化。
“告警恢复正常不能证明什么。在真实场景、负载压力下通过的测试才具备说服力。”
该服务还支持通过自然语言指令生成报表和仪表盘，内容可包含图表、表格、分页和链接，用于展示可靠性评分变化以及团队已处理的风险。这解决了可靠性团队的一个常见难题：被避免发生的事故无法像实际发生的故障那样产生直观的证据。
在另一篇文章中，Intellyx 分析师 Jason English 指出，可观测性和 AI SRE 产品有用，但大多数是被动型的。他认为，这些产品通常从已经发生问题的遥测数据开始，而故障注入则采取了不同的思路：在事故发生前创造受控的故障条件，并记录服务的响应。
English 还告诫说，可靠性计划不能简单地作为一个软件包来安装。要取得持续成效，仍然需要工程团队和管理层共同承担责任。这一点十分关键，因为 Foresight AI 可以协调测试与任务，但组织仍须决定哪些风险是可以接受的，以及是否采纳提议的变更方案。
此次发布也是行业将自动化弹性检测与交付流水线集成的众多实践之一，其中一些近期已被 InfoQ 报道。谷歌云混沌工程指南建议采用真实的故障条件、有限的故障影响范围以及通过 CI/CD 实现持续自动化。Wix 也描述了在 CI/CD 中使用 AI 进行日志解读和生成修复建议，同时将部署和关键基础设施决策权保留在人手中。
这些控制措施与 Foresight AI 的审批边界类似，同时又和高度自主化模型存在区别。例如，Cloudflare 的智能体开发生命周期（Agent Development Lifecycle）提案设想由智能体承担更多的测试、部署和维护工作。而 Gremlin 目前采取的是监督模式，使用智能体来提议测试和变更，必须经由人工授权才能执行。
在一个关于基于 AI SRE 自动化的 LinkedIn 讨论帖中，工程师 Sandip Bhattacharya 认为，企业仍然需要专家来配置智能体、通过指标评估其行为，并随着模型和基础设施的变化做出调整。这为 Gremlin 的监督模式提供了一个有用的检验：自动化或许能减少重复性工作，但也创造了一个可靠性工程师必须评估的新系统。
Amplify Partners 合伙人 Mike Dauber 将 Foresight AI 比作“替你完成重复训练的教练”。这个比喻说明了该产品对重复测试的侧重，尽管目前提供的证据来自 Gremlin 及其测试计划。不过 Gremlin 尚未公布对比准确率数据、误报率或详细的测试结果。
Gremlin 由曾任职于亚马逊和 Netflix 的工程师创立。 Andrus 此前曾在两家公司从事可靠性工程工作，Netflix 的 Chaos Monkey 帮助普及了自动化故障注入。Foresight AI 将这种方法从运行实验扩展到了测试筛选、结果解读、修复方案生成和持续验证。该产品经过内测验证，能够在风险演变为事故前识别并处理可靠性隐患，Gremlin Foresight AI 现已正式全面可用。
查看英文原文：https://www.infoq.com/news/2026/10/gremlin-foresight-ai/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Reliability management company Gremlin announced that its AI agent product Foresight AI was fully online. The product analyses potential failures in the service, gives recommendations for optimization and re-executs tests to verify the efficacy of the remediation programme. It is provided as an additional component of the Gremlin platform. The platform supports the team in conducting planned experiments, controlling the impact of failures and using automated tests to ensure long-term effectiveness of restoration measures. Enterprises can use the reliability rating quantization to improve progress and prioritize resource upgrading system elasticity.
Foresight AI aims to provide reliability detection capability without each developer becoming a distributed system specialist. Gremlin locates the product to respond to the quick delivery solution of AI support software. The CEO and co-founder of Gremlin Kolton Andrus argued that there was currently a large number of codes directly into the production environment without manual in-depth review.
“The best accident is never an accident.”
The product divides the work into four AI agent roles: the analyst AI agent (Analest) collects service data and recommends test examples; the tester AI agent (Tester) arranges and executes tests; the driver AI agent (Operator) interprets failed tests and proposes rehabilitation programmes; and the technical project manager AI agent (Technical Programme Manager) tracks test coverage, reliability risks and commitments of the teams. The last AI agent can generate weekly reports and send updates via Slack.
Foresight AI ' s recommendations are based on Gremlin ' s Fairure Atlas. According to the company, this is an exclusive repository of records containing millions of failed injection experiments carried out over the past 10 years on tens of thousands of distributed systems. Gremlin states that system data will not be used to train large language models. AI agent uses platform data and fault profiles for decision-making, while LLM is used only for searching, summarizing and interpreting results.
Its main technical feature is the closed-ring certification mechanism. When the tests reveal weaknesses in the system, the service proposes changes and reruns the original tests. Gremlin states that Foresight AI will not carry out any test or application restoration programme without obtaining permission. This keeps engineers responsible for changes and automates most of the analysis and validation work.
“There is nothing to prove that the police are back to normal. Tests passed under real scenes, load pressure are compelling.”
The service also supports the generation of reports and dashboards through natural language commands, which can contain charts, tables, page breaks and links to display reliability rating changes and the risks that the team has addressed. This solves a common difficulty for the reliability team: avoided accidents do not produce intuitive evidence like actual failures.
In another article, Intellyx analyst Jason English noted that observationability and AI SRE products were useful, but mostly passive. In his view, these products usually begin with telemetric data that have already caused problems, while failure injection takes different thinking: Create controlled malfunction conditions before an accident occurs and document the response of the service.
English also cautioned that the reliability plan could not be installed simply as a software package. Shared responsibility between the engineering team and management is still needed to achieve sustained results. This is critical as Foresight AI can coordinate tests and missions, but the organization still has to decide which risks are acceptable and whether to adopt the proposed changes.
This time publish is also one of the many practices of the industry that integrate automated elasticity testing with the delivery stream, some of which have recently been reported in InfoQ. The Google Cloud Dirt Project Guide recommends the introduction of real failure conditions, limited failure impact and continuous automation through CI/CD. Wix also describes the use of AI in the CI/CD for log reading and production of recommendations for restoration, while retaining the power to deploy and make decisions on critical infrastructure.
These controls are similar to the approved boundaries of Foresight AI, while at the same time distinguishing them from highly autonomous models. For example, the AI agent Development Life Cycle proposal for Claudflare envisages additional testing, deployment and maintenance by Z42Z. Gemlin, on the other hand, is currently following a monitoring model, using AI agent to propose tests and changes, which must be carried out with manual authorization.
In a LinkedIn discussion paper based on AI SRE automation, engineer Sandip Bhattacharya argued that the enterprise still needed experts to configure AI agent, assess its behaviour through indicators and adjust to changes in models and infrastructure. This provides a useful test for Gremlin ' s monitoring model: automation may reduce duplication of effort, but it also creates a new system that reliable engineers must assess.
Amplify Partners Partner Mike Dauber compares Foresight AI to "the coach who finishes your repeated training." This metaphor illustrates the focus of the product on repeated testing, although the evidence currently provided is from Gremlin and its testing plan. However, Gremlin has not published comparative accuracy data, misreporting or detailed test results.
Gremlin was created by engineers who worked in Amazon and Netflix. Andrews had previously worked in two companies on reliability engineering, and Chaos Monkey of Netflix had helped spread automated failure injections. Foresight AI has extended this method from operating experiments to test screening, interpretation of results, restoration programme generation and continuous validation. The product has been internally validated to identify and address reliability hazards before the risk evolves into an accident, and Gremlin Foresight AI is now fully available.
See the English original language at https://www.infoq.com/news/2026/10/gremlin-foresight-ai/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.